(74g) On the Development of a New Stochastic Optimization Algorithm with Capabilities for Distributed Computing

The paper presents the Cascade Optimization Algorithm, a distributed optimisation method featuring high-throughput computation capabilities and the potential to capitalize on knowledge management and knowledge acquisition applications without setting any computational burdens on the optimization search. The design of the algorithm bypasses the sequential nature of long Markov chains making use of a pool structure that stores intermediate solutions and providing means to coordinate the search. Pools feature populations redistributed periodically using a variety of acceptance criteria with multiple Markov processes launched simultaneously (high-throughput search) and following different partitions. In the course of the optimization search, populations can be analysed using different knowledge acquisition and discovery techniques (data mining, ontology engineering) further enabling the automated control of the algorithm.

The paper presents the mathematical foundation of the algorithm formalizing the cascade components, the cascade structure and its properties. The paper illustrates the performance of the algorithm with a variety of small and large-scale examples, mostly within reaction engineering. Results report the impact of the cascade structure, the distribution functions, the Markov length, and the different management strategies on convergence and acceleration. Results include comparisons with conventional methods, and selected implementations on parallel and distributed computing environments. Results indicate acceleration by 3.5 times with respect to a conventional stochastic search and, in addition, about 2.0 times with respect to conventional parallelization. The algorithm retains robustness in the solution of complex problems but, more importantly, it enables the enfranchised analysis of information available from intermediate solutions pointing to further extensions in the future.